Merge branch 'develop' into unsqueeze_op

Author: chenwhql

paddle/fluid/inference/analysis/data_flow_graph.cc

@@ -90,6 +90,20 @@ std::string DataFlowGraph::DotString() const {
   return dot.Build();
 }
 
+std::string DataFlowGraph::HumanReadableInfo(bool show_values,
+                                             bool show_functions) const {
+  std::stringstream values, functions;
+  for (auto &n : nodes.nodes()) {
+    if (show_values && n->IsValue()) {
+      values << n->repr() << "\n";
+    }
+    if (show_functions && n->IsFunction()) {
+      functions << n->repr() << "\n";
+    }
+  }
+  return "Values:\n" + values.str() + "\n\n" + "Functions:\n" + functions.str();
+}
+
 //
 // NodesBFSIterator
 //
@@ -146,7 +160,7 @@ bool GraphTraits<DataFlowGraph>::NodesBFSIterator::operator==(
   if ((!queue_.empty()) && (!other.queue_.empty())) {
     return queue_.front() == other.queue_.front() &&
            visited_.size() == other.visited_.size();  // here need to check the
-                                                      // equality of queue and
+    // equality of queue and
     // visited. Just a light but week implementation.
   }
   return false;
@@ -208,6 +222,76 @@ Node *GraphTraits<DataFlowGraph>::NodesDFSIterator::operator->() {
   return stack_.top();
 }
 
+GraphTraits<DataFlowGraph>::NodesTSIterator::NodesTSIterator(
+    const std::vector<Node *> &source) {
+  PADDLE_ENFORCE(!source.empty(),
+                 "Start points of topological sorting should not be empty!");
+  std::unordered_set<Node *> visited;
+  std::unordered_set<Node *> to_visit{source.begin(), source.end()};
+
+  std::vector<Node *> inlink_visited;
+  while (!to_visit.empty()) {
+    std::vector<Node *> queue(to_visit.begin(), to_visit.end());
+    for (auto *p : queue) {
+      inlink_visited.clear();
+
+      std::copy_if(p->inlinks.begin(), p->inlinks.end(),
+                   std::back_inserter(inlink_visited),
+                   [&](Node *x) { return visited.count(x); });
+
+      if (inlink_visited.size() == p->inlinks.size()) {
+        sorted_.push_back(p);
+        for (auto *_ : p->outlinks) {
+          if (!visited.count(_)) {
+            to_visit.insert(_);
+          }
+        }
+
+        to_visit.erase(p);
+        visited.insert(p);
+      }
+    }
+  }
+}
+
+GraphTraits<DataFlowGraph>::NodesTSIterator::NodesTSIterator(
+    const paddle::inference::analysis::GraphTraits<
+        DataFlowGraph>::NodesTSIterator &other)
+    : sorted_(other.sorted_), cursor_(other.cursor_) {}
+
+Node &GraphTraits<DataFlowGraph>::NodesTSIterator::operator*() {
+  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
+  return *sorted_[cursor_];
+}
+
+paddle::inference::analysis::GraphTraits<DataFlowGraph>::NodesTSIterator
+    &GraphTraits<DataFlowGraph>::NodesTSIterator::operator++() {
+  if (++cursor_ >= sorted_.size()) {
+    sorted_.clear();
+    cursor_ = 0;
+  }
+  return *this;
+}
+paddle::inference::analysis::GraphTraits<DataFlowGraph>::NodesTSIterator &
+GraphTraits<DataFlowGraph>::NodesTSIterator::operator=(
+    const paddle::inference::analysis::GraphTraits<
+        DataFlowGraph>::NodesTSIterator &other) {
+  cursor_ = other.cursor_;
+  sorted_ = other.sorted_;
+  return *this;
+}
+
+bool GraphTraits<DataFlowGraph>::NodesTSIterator::operator==(
+    const paddle::inference::analysis::GraphTraits<
+        DataFlowGraph>::NodesTSIterator &other) {
+  return sorted_ == other.sorted_ && cursor_ == other.cursor_;
+}
+
+Node *GraphTraits<DataFlowGraph>::NodesTSIterator::operator->() {
+  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
+  return sorted_[cursor_];
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/data_flow_graph.h

@@ -48,6 +48,9 @@ struct DataFlowGraph {
   // Output a DOT graph file for debug.
   std::string DotString() const;
 
+  std::string HumanReadableInfo(bool show_values = true,
+                                bool show_functions = true) const;
+
  private:
   // Remove duplicate edges and so on.
   void Clean();
@@ -107,6 +110,32 @@ struct GraphTraits<DataFlowGraph> {
     std::unordered_set<Node *> visited_;
   };
 
+  // Topological sorting iterator on nodes.
+  struct NodesTSIterator
+      : public std::iterator<std::forward_iterator_tag, Node *> {
+    NodesTSIterator() = default;
+    explicit NodesTSIterator(const std::vector<Node *> &source);
+    NodesTSIterator(NodesTSIterator &&other)
+        : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) {
+      other.cursor_ = 0;
+    }
+    NodesTSIterator(const NodesTSIterator &other);
+
+    Node &operator*();
+    NodesTSIterator &operator++();
+    // TODO(Superjomn) current implementation just compare the first
+    // element, need to compare the graph and all the elements in the queue and
+    // set.
+    NodesTSIterator &operator=(const NodesTSIterator &other);
+    bool operator==(const NodesTSIterator &other);
+    bool operator!=(const NodesTSIterator &other) { return !(*this == other); }
+    Node *operator->();
+
+   private:
+    std::vector<Node *> sorted_;
+    int cursor_{0};
+  };
+
   explicit GraphTraits(DataFlowGraph *graph) : graph_(graph) {}
 
   // default use BFS to visit the nodes.
@@ -119,17 +148,24 @@ struct GraphTraits<DataFlowGraph> {
   iterator_range<NodesDFSIterator> nodes_in_DFS() {
     return iterator_range<NodesDFSIterator>(nodes_dfs_begin(), nodes_dfs_end());
   }
+  iterator_range<NodesTSIterator> nodes_in_TS() {
+    return iterator_range<NodesTSIterator>(nodes_ts_begin(), nodes_ts_end());
+  }
 
  private:
   NodesBFSIterator nodes_bfs_begin() {
     return NodesBFSIterator(graph_->inputs);
   }
   NodesBFSIterator nodes_bfs_end() { return NodesBFSIterator(); }
+
   NodesDFSIterator nodes_dfs_begin() {
     return NodesDFSIterator(graph_->inputs);
   }
   NodesDFSIterator nodes_dfs_end() { return NodesDFSIterator(); }
 
+  NodesTSIterator nodes_ts_begin() { return NodesTSIterator(graph_->inputs); }
+  NodesTSIterator nodes_ts_end() { return NodesTSIterator(); }
+
  private:
   DataFlowGraph *graph_;
 };

paddle/fluid/inference/analysis/data_flow_graph_tester.cc

@@ -24,11 +24,11 @@ TEST(DataFlowGraph, BFS) {
   auto dfg = ProgramDescToDFG(desc);
   dfg.Build();
 
-  for (auto* in : dfg.inputs) {
+  for (auto *in : dfg.inputs) {
     LOG(INFO) << "inputs: " << in->name() << " "
               << static_cast<int>(in->type());
   }
-  for (auto* out : dfg.outputs) {
+  for (auto *out : dfg.outputs) {
     LOG(INFO) << "outputs: " << out->name() << " "
               << static_cast<int>(out->type());
   }
@@ -57,6 +57,71 @@ TEST(DataFlowGraph, DFS) {
   ASSERT_EQ(count, dfg.nodes.size());
 }
 
+// Topological sorting.
+/*
+ * Graph topology
+ * inputs: 0, 1, 2
+ * 0 -> 4
+ * 0 -> 5
+ * 1 -> 6
+ * 2 -> 7
+ * 4 -> 5
+ * 4 -> 7
+ * 4 -> 3
+ * 7 -> 3
+ */
+TEST(DataFlowGraph, TS) {
+  DataFlowGraph graph;
+
+  for (int i = 0; i < 8; i++) {
+    auto *node = graph.nodes.Create(Node::Type::kValue);
+    node->SetName("node-" + std::to_string(i));
+  }
+
+  auto add_link = [&](int i, int j) {
+    Node *source = graph.nodes.GetMutable(i);
+    Node *target = graph.nodes.GetMutable(j);
+    target->inlinks.push_back(source);
+    source->outlinks.push_back(target);
+  };
+
+  graph.inputs.push_back(graph.nodes.GetMutable(0));
+  graph.inputs.push_back(graph.nodes.GetMutable(1));
+  graph.inputs.push_back(graph.nodes.GetMutable(2));
+
+  add_link(0, 4);
+  add_link(0, 5);
+  add_link(1, 6);
+  add_link(2, 7);
+  add_link(4, 5);
+  add_link(4, 7);
+  add_link(4, 3);
+  add_link(7, 3);
+
+  auto its = GraphTraits<DataFlowGraph>(&graph).nodes_in_TS();
+  std::vector<int> sorted_ids;
+  for (auto it = its.begin(); it != its.end(); ++it) {
+    LOG(INFO) << it->name();
+    sorted_ids.push_back(it->id());
+  }
+
+  // Assert a occurs prior to b in the sorted_ids.
+  auto assert_positive_sequence_pair = [&](int a, int b) {
+    auto a_offset = std::find(sorted_ids.begin(), sorted_ids.end(), a);
+    auto b_offset = std::find(sorted_ids.begin(), sorted_ids.end(), b);
+    ASSERT_LT(a_offset, b_offset);
+  };
+
+  assert_positive_sequence_pair(2, 7);
+  assert_positive_sequence_pair(7, 3);
+  assert_positive_sequence_pair(4, 3);
+  assert_positive_sequence_pair(0, 4);
+  assert_positive_sequence_pair(0, 5);
+  assert_positive_sequence_pair(1, 6);
+  assert_positive_sequence_pair(4, 5);
+  assert_positive_sequence_pair(4, 7);
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/operators/CMakeLists.txt

@@ -266,6 +266,7 @@ op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale)
 op_library(cos_sim_op DEPS cos_sim_functor)
 op_library(parallel_do_op DEPS executor)
 op_library(unsqueeze_op DEPS reshape_op)
+op_library(squeeze_op DEPS reshape_op)
 
 if (WITH_GPU)
     op_library(conv_op DEPS vol2col depthwise_conv im2col)

paddle/fluid/operators/detection/rpn_target_assign_op.cc

@@ -86,8 +86,9 @@ class RpnTargetAssignKernel : public framework::OpKernel<T> {
                          std::minstd_rand engine,
                          std::vector<int>* inds) const {
     std::uniform_real_distribution<float> uniform(0, 1);
-    if (inds->size() > num) {
-      for (int i = num; i < inds->size(); ++i) {
+    const int64_t size = static_cast<int64_t>(inds->size());
+    if (size > num) {
+      for (int64_t i = num; i < size; ++i) {
         int rng_ind = std::floor(uniform(engine) * i);
         if (rng_ind < num)
           std::iter_swap(inds->begin() + rng_ind + offset,

paddle/fluid/operators/im2sequence_op.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/im2sequence_op.h"
+#include <string>
 #include <vector>
 
 namespace paddle {
@@ -28,20 +29,19 @@ class Im2SequenceOp : public framework::OperatorWithKernel {
                    "Input(X) of Im2SequenceOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of Im2SequenceOp op should not be null.");
-
     auto in_dim = ctx->GetInputDim("X");
+
     PADDLE_ENFORCE_EQ(in_dim.size(), 4,
                       "Input(X) format must be 4D tensor, eg., NCHW.");
-
-    auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
-    auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
-    auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
-
     int batch_size = in_dim[0];
     int img_channels = in_dim[1];
     int img_height = in_dim[2];
     int img_width = in_dim[3];
 
+    auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
+    auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
+    auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+
     int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0],
                                          paddings[2], strides[0]);
     int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1],
@@ -61,6 +61,10 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
              "C: channels"
              "H: height"
              "W: width");
+    AddInput("Y",
+             "(Tensor) The input tensor of image real size(H, W)."
+             "2-D with shape [batchsize, 2]")
+        .AsDispensable();
     AddOutput("Out", "(LodTensor) The output data of im2sequence op,");
     AddAttr<std::vector<int>>("kernels",
                               "(vector<int>), the "
@@ -73,6 +77,13 @@ class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
                               "(vector<int> default:{0, 0, 0, 0}), the "
                               "paddings(up_pad, left_pad, down_pad, right_pad)")
         .SetDefault({0, 0, 0, 0});
+    AddAttr<std::vector<int>>("out_stride",
+                              "the attribute is valid only when input(Y)"
+                              "is not NULL.this attribute represents the"
+                              "scaling of the pic through the CNN"
+                              "(vector<int> dedault:{1,1}),the out_stride"
+                              " (out_stride_height, out_stride_width)")
+        .SetDefault({1, 1});
     AddComment(R"DOC(
 This op uses kernels to scan images and converts these images to sequences.
 After expanding, The number of time steps are output_height * output_width
@@ -123,7 +134,7 @@ output.data = [[ 6.  2.  8.  3.  2.  4.  6.  3.]
                [ 7.  1.  7.  9.  2.  1.  3.  5.]
                [ 5.  7.  2.  4.  1.  3.  9.  0.]
                [ 7.  9.  4.  8.  3.  5.  0.  8.]]
-output.dims = {8, 9}
+output.dims = {8, 8}
 output.lod = [[0, 4, 8]]
 
 )DOC");